Unlocking New Insights with Artificial Intelligence
The past year has seen rapid growth of artificial intelligence (AI) technology across a wide range of fields, and ISPOR International 2023 showed that HEOR is no exception. While AI was the focus of the second plenary session,1 we learned throughout the conference about the many ways in which AI stands to innovate and disrupt the field. Two key use cases for AI stood out for having particular potential to revolutionize HEOR. First, AI enables the rapid analysis of large datasets. The amount of patient health data generated on a daily basis continues to grow through advances in wearable technology, digital therapeutics, and other innovations; with AI tools, insights from these datasets become increasingly accessible. Second, AI enables the creation of new, large datasets from existing unstructured data, increasing their utility. Vast amounts of data already exist in formats that are difficult or impossible to analyse, such as chart notes or qualitative patient-reported data; with AI tools, these data can be coded into structured datasets for easy analysis.
Using AI to Facilitate Data Analysis
Digital health technologies such as wearable sensors enable the collection of vast amounts of patient health data, but these data are only useful if they can be adequately analysed. One session discussed the role of AI in the analysis of clinical data from wearable devices in Duchenne’s muscular dystrophy (DMD).2 Using a machine learning model, data from the wearable sensors were processed to identify clinically meaningful disease features. After the model was adequately trained, it was able to predict disease progression far earlier than the traditional prognostic methods, including clinician scoring of a standardised 6-minute walking test. With sufficient validation, these technologies could be scaled up and used to define novel endpoints for DMD clinical trials, which could lead to reductions in the length of time and sample size necessary for a trial.2
The ability to rapidly analyse large datasets would impact more than trial design – research presented at the conference also highlighted the potential for AI and machine learning models in conducting literature reviews, which could drastically reduce the amount of time, effort and resulting costs associated with conducting rigorous reviews.3-5 While AI models are not infallible, they were shown to reduce overall time spent when working in conjunction with a human reviewer as part of a dual-review process.3
In the short term, overcoming inefficiencies with AI would save time and reduce costs, but the incorporation of AI into literature reviews also creates further opportunities. One poster highlighted the potential for living reviews to remain constantly updated with the help of AI for review, extraction and text generation.6 While these technologies would require sufficient validation before use in high-stakes scenarios, accessing data is likely to become increasingly easy with the help of AI tools.
Using AI to Create Large, Structured Datasets from Currently Unstructured Data
In addition to the increased accessibility of existing datasets, the use of AI technologies can expand the utility of data that are already available but difficult to access. One challenge associated with real world evidence generation is that key clinical data can be difficult to access and analyse systematically, such as physician notes on a patient chart. To extract useful evidence from physician notes, substantial manual effort is required to review and translate notes into an easily analysed data set. As discussed in the second plenary session, a large language model may be able to rapidly dissect physician notes from patient charts, reducing the timelines for a chart review from months to days.1
Going one step further, the second plenary session also highlighted the potential for AI as ambient technology, or in other words, a technology that would operate unnoticed in the background.1 In the case of clinical interactions, this could take the form of AI generating a patient chart automatically based on audio from a doctor-patient interaction, opening the door for easy analysis and reducing the paperwork burden for providers.
Taken together, these and other advances in AI enable far-reaching innovation across the field of HEOR, but seeing their impact may take time. Regulatory bodies including the FDA were well represented in both the second plenary session and in related issue panels, and representatives shared their support for the use of AI technologies in healthcare.1 However, while regulatory encouragement is beneficial, the limitations and risks of these new technologies must also be considered.
Risks and Limitations
Across AI-related sessions and research, experts emphasised that a degree of caution must be exercised during the implementation of AI technologies to minimise associated risks. AI technologies in particular are vulnerable to issues with data protection and privacy, technological limitations, and the potential to further contribute to health disparities for patients whose profile may not be well represented in the data used to train a given AI technology. However, there was an overall sense of optimism that the benefits of responsibly implemented AI are likely to outweigh the risks.
Applications for AI in HEOR are still in their very early stages, and there is certainly more innovation to come. Costello Medical’s in-house Technology Development team have established an AI task force to investigate the potential applications for AI in our work. While we do not yet use AI tools substantively or routinely due to the limitations outlined above and a lack of clarity surrounding the value they currently bring versus conventional approaches, we look forward to engaging with these technologies as they continue to evolve.
